WTF is Weight-space Programming?
AI agent jargon, decoded. Plain-English and technical explanations of the concepts behind AI agents, with sources.
All Jargons newest first
Weight-space Programming
An LLM reads and directly edits another neural network's weights to meet a plain-English goal, with no training run.
Agent Instruction Bloat
Agent instruction files that grow long, stale and contradictory until they make the agent worse.
Human-in-the-loop Theater
A human approval step that looks like oversight but is really rubber-stamping things they can't meaningfully check.
Stateful Guardrails
Guardrails that judge the whole conversation, not just the latest message, and remember strikes across a session.
Agentjacking
Hijacking a coding agent by planting instructions in data it already trusts, such as error logs it reads over MCP.
LLM Wiki
Instead of retrieving from raw documents, an agent builds and maintains its own structured wiki of what it has learned.
Agentic Flywheel
Agents improving their own harness, using feedback from tests, pipelines and production to recommend changes.
Human On The Loop
Instead of approving every agent output, humans shape the harness that governs how agents work.
Claw
A personal, always-on agent that runs on your own hardware, takes requests over chat apps, and can schedule its own tasks.
Harness Engineering
Improving everything around the model (tools, config, checks, AGENTS.md) so the agent stops making the same mistake.
Agentic Engineering
Professional software development with coding agents, where the developer still specifies, verifies and owns the result.
Agent Skills
Folders of instructions, scripts and resources that an agent loads only when a task needs them.
Context Rot
Models get measurably worse as their context grows, even when the extra tokens are irrelevant.
Ralph Wiggum Loop
Running a coding agent in a dumb bash loop, with a fresh context each time, until the spec is satisfied.
Agentic Misalignment
When an agent deliberately chooses a harmful action, like blackmail, because it's the only way left to reach its goal.
Lethal Trifecta
The dangerous combination of private data access, exposure to untrusted content, and a way to send data out.
Agent Washing
Rebranding chatbots, assistants or plain automation as "AI agents" when they have little real autonomy.
Context Engineering
Deciding exactly what goes into a model's context window at each step: instructions, tools, data and history.
Subagents
Delegating a narrow task to a helper agent with its own context, so only a summary comes back to the main agent.
Asynchronous Coding Agent
A fire-and-forget coding agent that works in its own hosted environment and comes back with a pull request.
A2A
An open protocol for AI agents built by different vendors to discover each other and hand off work.
Slopsquatting
Registering a malicious package under a name that AI coding tools predictably hallucinate.
Tool Poisoning
Hiding malicious instructions inside a tool's description so any agent that loads the tool is compromised.
Vibe Coding
Building software by describing what you want to an AI and accepting its code without closely reading it.
Compaction
When a session nears the context limit, the agent replaces older turns with a summary so it can keep going.
Spec-driven Development
Writing a structured spec before asking an AI to code, so the spec becomes the source of truth for both human and agent.
Orchestrator–workers
A lead model breaks a task into pieces, hands them to worker models, then combines their results.
Alignment Faking
A model behaves as trained when it thinks it's being watched, while planning to act differently when it isn't.
MCP
An open protocol that lets AI apps plug into external tools and data through a standard interface.
Computer Use
Agents that operate a computer like a person: looking at the screen, moving the mouse and typing.
Agentic Workflow
LLM calls wired together along a path you define in code, rather than letting the model choose every step.
AI Engineer
A software engineer who builds products on top of foundation models, rather than training models.
LLM-as-a-judge
Using one LLM to grade or compare the outputs of another against a rubric.
Agentic Loop
The core cycle of an agent: call the model, run the tools it asks for, feed results back, repeat until done.
Agent Memory
Ways for an agent to remember information beyond a single context window or conversation.
Reflection
An agent reviews its own output or failed attempt, writes down what went wrong, and tries again.
Multi-agent System
Several LLM agents, often with different roles or tools, working together on one task.
Tool Use
A model's ability to call external functions (APIs, code, search) instead of only producing text.
Evals
Test suites that measure how well an AI system performs on the tasks you actually care about.
Grounding
Tying a model's answer to verifiable sources so it states facts it can back up instead of guessing.
Guardrails
Checks around an AI system that block, filter or fix unsafe or invalid inputs and outputs.
ReAct
A pattern where the model alternates between thinking out loud and taking an action, then observes the result.
Prompt Injection
Malicious instructions hidden in content an AI reads, tricking it into ignoring its real instructions.
Chain-of-thought
Getting a model to write out intermediate reasoning steps before giving its final answer.
RAG
Fetching relevant documents first and putting them in the prompt, so the model answers from real data.
Prompt Engineering
Iteratively designing the instructions and examples you give a model to get reliable, useful outputs.
Embeddings
Lists of numbers that represent the meaning of text (or images) so similar things end up close together.
Blackboard Pattern
Agents coordinate by reading and writing to a shared space instead of messaging each other or reporting to a boss.
Agent Sandboxing
Running agents in isolated environments with restricted privileges, so mistakes or attacks can't reach beyond the box.
Context Window
The maximum number of tokens a model can take into account in one request, prompt and response included.
Fine-tuning
Further training a pretrained model on your own examples to specialise its behaviour.
Human-in-the-loop
Pausing an agent so a person can approve, correct or reject a step before it continues.
Time To First Token
How long a model takes to start responding, the latency metric that most shapes how fast an AI feels.
Tokens
The chunks of text (words, word pieces or characters) a model actually reads and writes, and what you pay for.
No jargons match. If it's new, the pipeline will probably pick it up soon.